**The most surprising finding is that the dataset is not a neutral record of tennis. Each player's own rank is systematically better than the rank of the opponents they face.** In `raw.raw_kaggle`, every match should appear once from each player's side. Across the whole table, average player rank (`Rk`) should therefore roughly equal average opponent rank (`vRk`). They don't match on any surface. This comes from the per-surface gold table (`raw_kaggle_by_Surface`), where a lower rank number means a better player: | Surface | Avg player rank (Rk) | Avg opponent rank (vRk) | |---|---|---| | Hard | 310.8 | 378.6 | | Clay | 361.4 | 433.7 | | Grass | 151.2 | 176.2 | | Carpet | 517.4 | 604.7 | On every surface, opponents are ranked about 15–20% worse than the tracked players. The medians show the same gap, for example Hard 237 vs 251 and Clay 288 vs 313. My reading is that the table is built around a fixed set of players, and their opponents are often outside that set. The table has only about 491 distinct `Name` values but 237,205 rows, so many opponents never get a row of their own. This is an inference from the asymmetry, not something I tested directly. Three other counterintuitive points came up in the same evidence: - **Large gaps in match statistics:** 86,793 of 237,205 rows (about 37%) have no serve statistics (`TP`, `Aces`, `DFs`, `SP`, `1SP`, `2SP`, `vA` are all null). Any analysis of serve performance silently drops more than a third of matches. - **The dataset is mostly lower-tier tennis:** The average player rank is 311 on hard courts and 361 on clay. The data is dominated by Challenger-level players, not the top of the ATP tour, despite the "tennis player data" framing. Grass is the exception, with a much better average rank of 151. - **Surface effects are big but unsurprising:** Average aces per match are 7.6 on grass and 7.9 on carpet versus 3.4 on clay, about 2.2 times as many on grass as on clay. **Implication:** Win-probability or scouting models built on `Rk` vs `vRk` would inherit this selection bias. Raw averages of serve and return stats would also not generalise to all players.